Research map / Large language models / Agents, code and math
LogicSAGE: Neuro-Symbolic Reasoning with Socratic-Guided Enhancement
Jinlong Tian, Jiang Yu, Kewei Cheng, Fengxiang Cheng, Yue He, Yunfei Wang, Haotian Wang, Haoxuan Li, Wenjing Yang, Shixuan Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reactive error-guided correction (ablation) | LogicSAGE (SFT Qwen3-8B) · 5 logic benchmarks avg | execution accuracy: 70.1 -> 88.4 | Improved |
| add reactive correction (ablation) | LogicSAGE (SFT Qwen3-8B) · AR-LSAT | execution rate: +63.6 | Improved |
| add proactive uncertainty probing on top of reactive (ablation) | LogicSAGE (SFT Qwen3-8B) · 5 logic benchmarks avg | execution accuracy: 88.4 -> 91.2 | Improved |
| add proactive probing (ablation) | LogicSAGE (SFT Qwen3-8B) · AR-LSAT | execution accuracy: 66.5 -> 74.6 | Improved |
| add proactive probing (ablation) | LogicSAGE (SFT Qwen3-8B) · PrOntoQA, ProofWriter execution rate | execution rate: +0.0 | No clear effect |
| add symbolic System 2 to neural System 1 (fusion) (ablation) | LogicSAGE · 5 logic benchmarks avg | accuracy: 80.18 -> 92.36 | Improved |
| add neural fallback and arbitration to symbolic solver (ablation) | LogicSAGE · 5 logic benchmarks avg | accuracy: 90.54 -> 92.36 | Improved |
| add neural fallback and arbitration (ablation) | LogicSAGE · AR-LSAT | accuracy: 71.32 -> 77.00 | Improved |
| supervised fine-tuning on solver-verified data | LogicSAGE (Qwen3-8B) · 5 logic benchmarks avg | average accuracy: 80.30 -> 92.36 | Improved |
| increase refinement depth Tmax from 0 to 3 (ablation) | LogicSAGE · 5 logic benchmarks avg | average accuracy: 79.30% -> 92.36% | Improved |
| increase refinement depth Tmax from 3 to 4 (ablation) | LogicSAGE · 5 logic benchmarks avg | average accuracy: < 0.2% | No clear effect |
| swap backbone from Qwen3-8B (training-free) to GPT-4 | LogicSAGE · 5 logic benchmarks avg | average accuracy: 80.30 -> 86.10 | Improved |
Closest papers by meaning
- Logically Consistent Language Models via Neuro-Symbolic Integration (ICLR 2025 · similarity 0.89)
- Divide and Translate: Compositional First-Order Logic Translation and Verification for Complex Logical Reasoning (ICLR 2025 · similarity 0.89)
- Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment (ICML 2025 · similarity 0.88)
- Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI (NeurIPS 2025 · similarity 0.88)
- NeSTR: A Neuro-Symbolic Abductive Framework for Temporal Reasoning in Large Language Models (AAAI 2026 · similarity 0.88)
- Large Language Models Meet Symbolic Provers for Logical Reasoning Evaluation (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
